Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
RuleViz: A Model for Visualizing Knowledge Discovery Process Jianchao Han Nick Cercone Department of Computer Science University of Waterloo 200 University Avenue West Waterloo, Ontario, N2L 3G1 Canada Department of Computer Science University of Waterloo 200 University Avenue West Waterloo, Ontario, N2L 3G1 Canada j2han@hopper.uwaterloo.ca ncercone@math.uwaterloo.ca ABSTRACT Categories and Subject Descriptors I.5.5 [Pattern recognition]: Implementation|Interactive systems; I.2.6 [Articial intelligence]: Learning|Induction, Concept learning; H.2.8 [Database Managemen t]: Database Applications|Data Mining; H.5.m [Information interfaces and presentation]: Miscellaneous We propose an interactive model, RuleViz, for visualizing the entire process of knowledge discovery and data mining. The model consists of ve components according to the main ingredients of the knowledge discovery process: original data visualization, visual data reduction, visual data preprocess, visual rule discovery, and rule visualization. The RuleViz model for visualizing the process of knowledge discovery is introduced and each component is discussed. Two aspects are emphasized, human-machine interaction and process visualization. The interaction helps the KDD system navigate through the enormous search spaces and recognize the intentions of the user, and the visualization of the KDD process helps users gain better insight into the multidimensional data, understand the intermediate results, and interpret the discovered patterns. General Terms Interactive exploration, data and knowledge visualization, knowledge discovery, rule induction. 1. INTRODUCTION Knowledge discovery in databases (KDD) is a process of discovering useful patterns from the original large data sets. The KDD process is interactive and iterative, involving numerous steps with many decisions being made by the user [13]. Most existing methods for knowledge discovery reported in the literature stresses the interactions between the human and the machine. Their implementations, however, focus on the logical algorithms for data mining rather than the interactions. Some visualization systems develop either the original data visualization or the nal results visualization, but few of them emphasize the visualization of the entire process of KDD and the interactions for navigating through the search space. According to the RuleViz model, we implement an interactive system, CViz, which exploits \parallel coordinates" technique to visualize the process of rule induction. The original data is visualized on the parallel coordinates, and can be interactively reduced both horizontally and vertically. Three approaches for discretizing numerical attributes are provided in the visual data preprocessing. CViz learns classication rules on the basis of a rule induction algorithm and presents the result as the algorithm proceeds. The discovered rules are nally visualized on the parallel coordinates with each rule being displayed as a directed \polygon", and the rule accuracy and quality are used to render the \polygons" and control the choice of rules to be displayed to avoid clutter. The CViz system has been experimented with the UCI data sets and synthesis data sets, and the results demonstrate that the RuleViz model and the implemented visualization system are useful and helpful for understanding the process of knowledge discovery and interpreting the nal results. Currently, two important approaches for knowledge discovery can be characterized as algorithm-based and visualizationbased [8]. The former includes rule induction [3], concept learning [27], association mining [26], decision tree induction [30], neural networks [10], etc. These methods specify the target formats and pursue the associations between the outcome variable and the independent variables. On the other hand, the latter species the hypothesis by means of visualization metaphors. Many techniques and systems for data visualization and/or knowledge visualization have been developed and implemented [2, 5, 7, 11, 12, 17, 18, 23, 24]. These visualization systems can be divided into two categories, one visualizing the raw data like Spotre [2], Independence Diagrams [7], Periodic Data [11], DViz [18], and VisDb [23], and the other visualizing the nal results like Decision Tree Construction [5], Map Visualizer and Tree Visualizer developed by Silicon Graphics [17]. One common feature of these existing systems is their dependence on computer graphics and scientic visualization. Most visualization systems lack the ability to visualize the entire pro- Permission to make digital or hard copies of part or all of this work or personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or a fee. KDD 2000, Boston, MA USA © ACM 2000 1-58113-233-6/00/08 ...$5.00 244 fully automated process of knowledge discovery because this requires that the system possess all domain knowledge and be capable of recognizing the intentions of the user. Hence, eective human-machine interaction plays an important rule in the KDD process. Various visualization tools can be used to concern the interaction and help incorporate the user's intentions and navigate through the enormous search spaces interactively. cess of knowledge discovery. In data visualization systems, the complex data are carefully arranged and displayed in a specic visual form, and the knowledge behind the original data is left to the users who must observe and determine the meaning of the pictures. This usually requires a wealth of background knowledge on graphics and applications. On the other hand, the knowledge visualization systems visualize the discovered knowledge according to dierent techniques such as decision trees, neural networks, etc. However, these systems only display the results on the basis of some parameters specied by the user, and the user can not understand how the results are obtained and interpreted. Interactive visualization system of knowledge discovery should provide a user with not only the original data and/or the discovered results but also the entire process in visual forms so that the user can participate in the discovery process, provide heuristics, guide the navigation of search, and interpret the discovered knowledge. We propose an interactive visualization model, RuleViz, which stresses two aspects, human-machine interaction and visualization during the entire KDD process. The interaction between the user and the machine helps the KDD system navigate through the enormous search spaces and recognize the intentions of the user, and the user easily provides the system with heuristics and domain knowledge and species parameters. On the other hand, the visualization of the KDD process helps users gain better insight into the multidimensional data, understand the intermediate results, and interpret the discovered patterns. We also implement an interactive system, CViz, according to the RuleViz model for visualizing the process of classication rule induction. The CViz system uses parallel coordinates technique [22, 24] to visualize the original data and the discretized data, and the discovered rules are also visualized as rule polygons on the parallel coordinates system with dierent colors representing the rule accuracy and the rule quality. 2.1 Original Data Visualization Original data visualization is to use a variety of data visualization techniques to appropriately display the original data in some visual form. How the original data is visualized depends on the patterns that are to be discovered. The goal of this component is to give the user insight that is virtually impossible to get from looking at tables of output or simple summary statistics. The original data visualization can tell how the raw data is distributed, where the interesting data focuses on, and whether the patterns can be observed. The patterns to be searched for might be classication rules or association rules, at rules or hierarchical rules, etc. Dierent patterns may require dierent visual forms in order for the user to observe and determine the interesting areas. For example, classication rules may require that the focus be on the relationships between the condition attributes and their values and decision attribute and their class labels. The association rules may require the focus to be on the relationships among item subsets for item transaction data sets, or the relationships among the value ranges of dierent attributes for numerical data sets. The decision trees may require the focus to be on the hierarchy of concepts or the distribution of data in dierent regions (branches). This paper is organized as follows. The RuleViz model and its ve components are proposed in Section 2. In Section 3, the CViz system is presented, and the experiments with CViz are illustrated. Finally, Section 4 is the concluding remarks. 2. In order to construct such interactive systems, we suggest an interactive visualization model, RuleViz, to visualize the entire process of knowledge discovery. It consists of the ve components: original data visualization, interactive data reduction (horizontally and vertically), visual data preprocess such as numerical attribute discretization and data transformation, pattern discovery like rule induction and decision tree construction, and pattern visualization, shown in Figure 1. According to this model, we implement an interactive visualization system, CViz, for rule induction. In this section, we briey describe each component and explain what techniques can be used in each step. THE RULEVIZ MODEL According to [24], the goals of data visualization could be one of three aspects: exploratory analysis, which searches for the structures, trends, and clusters of the data; conrmatory analysis, which conrms or rejects the pre-dened hypotheses; and presentation, which presents the facts. In our model, we assume the original data visualization is to present the data. Generally speaking, the knowledge discovery in databases is considered to be the nontrivial process of identifying valid, novel, potentially useful, and ultimately understandable patterns in data [14]. It uses the database along with any required selection, preprocessing, sub-sampling, and transformations to apply data mining algorithms to enumerate patterns and to evaluate the products of data mining to identify the subset of the enumerated patterns deemed \knowledge" [10]. This process underlines an interactive and user-friendly character of overall activities. Many discussions about the KDD process and its main ingredients are present in the literature. Basically, the KDD process consists of data collection, target data creation, data cleaning, data preprocessing, data reduction, data mining, pattern interpretation, and knowledge application [8]. The KDD process relies on a human-system interaction. It is impossible to envision a There are many data visualization techniques that can be used in our model, including geometric techniques, iconbased techniques, pixel-oriented techniques, hierarchical techniques, graph-based techniques, dynamic techniques, and their combination. It must be emphasized that the distance measure between data points when visualized is so important that inappropriate measure will lead the user in a wrong direction or into a mistaken action in the follow- 245 Data Visualization Routines Original Data Visualization Original Data Set Data Reduction Methods Data Preprocess Approaches Algorithms for Pattern Discovery Pattern Visualization Routines Data Preprocess Pattern Discovery Pattern Visualization Data Reduction Reduced Data Set Preprocessed Data Set Discovered Patterns Figure 1: The RuleViz Model ing steps. Usually, the distance measure can be Euclidean, Mankowski, Tschebyschev, and others. they think less interesting by clicking the corresponding coordinates. In the CViz system [20], the parallel coordinates approach, a kind of geometric technique, is used to visualize the original data, where each hyperpoint is displayed as a polyline across the parallel coordinates. The other category addresses example selection, or tuple selection (horizontal reduction) which selects or searches for a representative portion of the original data set that can fulll a knowledge discovering task as if the whole data set is exploited [10, 21]. Example selection techniques can be divided into two main classes: parametric techniques and nonparametric techniques [6]. The former assumes a parametric model for the data distribution, like Gaussian or exponential distributions, and then estimates the parameters of the model. After the parameter estimate is obtained the typical data, with respect to the parametric model, can be drawn according to their typicality measure, which forms the representative portion of the original data [21]. The latter does not assume any model for the data. The original data can be randomly sampled or sampled based on the use of criterion functions. 2.2 Data Reduction Data reduction is to reduce the redundant or insignicant data, or select a small portion to represent the entire set due to time complexity and/or main memory capacity. This is necessary for real-world applications as databases are usually huge. For example, the database in Wal-Mart is increased over 20 million transactions a day [13]. However, many knowledge discovery techniques are very sensitive to the size of data in terms of time complexity and storage complexity. For interactive knowledge discovery, the original data can be cleaned by the user who may observe the data distribution through the visual form of the original data and determine which data area he/she is interested in, which attributes are more important than others, and which attribute values are redundant and can be removed. The numerical attribute discretization and concept hierarchy climbing can also be used to reduce the original data set horizontally. By discretizing the numerical attributes into interval attributes, the tuples that fall in the same hyperintervals can be combined. In the RuleViz model, numerical attribute discretization is done in the next component, data preprocessing. By climbing the concept hierarchy, feature values can be substituted by their higher level values in the predened hierarchy of categorical values of this attribute. In CViz, users can remove the uninteresting attribute values by clicking corresponding coordinates portion, thus all tuples that pass by this attribute-value point will be removed from the data set. Many approaches to data reduction have been proposed and implemented for knowledge discovery [21]. Basically, two categories of algorithms have been investigated and developed extensively. One category focuses on the reduction of features or attributes (vertical reduction) which are used to represent the original data, called feature selection [10, 25]. Feature selection is trying to nd the best feature subset from the original feature set according to a given selection criterion. The goal is to reduce the data dimensionality by selecting important attributes without or with less loss of information behind the data so that the data can be expressed in some succinct formation and the data size is decreased to the extent that the algorithms can eÆciently execute in the main memory. In CViz, users can remove attributes which The outcome of data reduction is a subset of the original data set which has been cleaned vertically and horizontally, and will be used to be the representative of the entire data set. This step can be repeated until the selected data subset 246 ber of intervals such that each interval contains the roughly same number of sorted attribute values; entropy-based partition [4, 10, 30], which discretizes the attributes based on maximal entropy of the partition; and distance-based partition [26], which partitions the attribute value range by identifying and combining clusters of data points. is satisfactory in terms of the user's intentions or some prespecied criteria. 2.3 Data Preprocess Data preprocess is a sequence of operations that convert the original data to the target format for the data mining algorithms. It includes several aspects such as data normalization, data transformation and feature extraction, handling the missing attribute values and noisy data, numerical attribute discretization, and so on [10]. In CViz, three approaches for discretizing numerical attributes are provided, including equi-sized partition, bin-packing based equi-depth approach, and the EDA-DB (Entropy-based Discretization According to Distribution of Boundary points) method [20], where the bin-packing based approach is an image-based approach, while the other two are algorithmbased approaches. When the data is visualized or processed for the knowledge discovery, they may require scaling and normalization. A simple normalization formula is: x x xnormalized = ; x where x and x are the mean and standard deviation of attribute x, respectively. 2.4 Pattern Discovery Pattern discovery is the task of data mining, an important step of KDD, which searches for patterns of interest in a particular representation form. Data transformation is to transform the current data into appropriate format which can be used directly by the specic knowledge discovery algorithms. For example, in the CViz system, ELEM2 is chosen as the rule induction algorithm, which requires that the data and the nal rules be in the form of < attribute; value > pairs. Another transformation is feature extraction, which may lower the dimensionality of the patterns by extracting the most informative data features based on the attribute dependency analysis. It transforms the raw data into condensed representation, resulting in the reduction of data. Principle coordinates analysis (PCA) is an important approach for extracting features [16]. Basically, for p-dimensional space, by analyzing the distance matrix between data points, PCA searches for s < p and denes s axes to approximately represent the original space. There have been two approaches for pattern discovery so far. One is algorithm-based, and the other is image-based. Algorithm-based approach stresses the logic and/or statistics reasoning and attempts to nd the patterns behind the data mathematically. Decision tree techniques [30], neural networks, statistics methods [27] association rule mining algorithms [26], rule induction [3], and case-based reasoning [27] belong to this category. Several solutions to the problem of generating decision tree or decision rules from the training set of examples with unknown values have been proposed [27, 30]. The simplest among them consists in removing examples with unknown values or replacing missing values with the most common values. More complex approaches are to use a Bayesian formalism to determine the probability distribution of the missing value over the possible values from the domain, and either choose the most likely value or divide the example into fractional examples, each with one possible value weighted according to the probabilities determined. In our implementation, CViz embeds a learning algorithm ELEM2 [3] to discover classication rules, while AViz is an image-based approach for discovering numerical association rules for large data sets [19]. The other category of approaches for nding patterns from large data sets is to directly use the images that visualize the preprocessed data [19, 15]. The characteristics such as shape, color, brightness, etc. of the image provides us diverse information which can be developed to discover patterns. This approach depends on the visualization techniques. 2.5 Pattern Visualization The last component is pattern visualization which is to visualize the patterns achieved in pattern discovery on the screen. Since dierent patterns may have dierent characteristics, they must be drawn with dierent techniques. A collection of visualization routines should be developed to adapt to dierent patterns to reect their characteristics. For example, a tree structure can be used to represent hierarchical concepts. Icon shapes and colors which change with sliders can be used to represent the associations between attributes. Animation and uid ow can be used to show the evolution of attributes. As mentioned previously, one important task of data preprocess is to discretize numerical attributes. Most approaches for discovering knowledge discretize the numerical attributes into intervals and then treat the discretized attributes as categorical ones and intervals as categorical values. A discretization process consists of two general steps. The rst is deciding the number of discrete intervals, and the second is determining the boundaries of each intervals. There are many approaches to this problem, such as equi-sized partition [15, 18], which nds the minimum and maximum values for each numerical attributes and divides this range into a number of user-specied equal width discrete intervals; equidepth partition [29], which sorts the values of each attribute and then divides these values into the user-specied num- There have been several pattern visualization systems. Silicon Graphics developed a data mining package, MineSet, which has several visualization components [9, 17], including Map Visualizer and Tree Visualizer. The Map Visualizer is used to display quantitative and relational characteristics of spatially oriented data, where data items are associated with graphics \bar chart" objects in the visual landscape which can consist of a collection of spatially related objects, each 247 the knowledge discovery can be interactively removed by \deleting-click", thus the data is vertically reduced. While an attribute value can be deleted if the distribution of the data tuples passing the value is too sparse, thus the data tuples that have this value could be removed so that the data is horizontally reduced. The data reduction is highly dependent on the user [18]. with individual heights and colors. The Tree Visualizer displays quantitative and relational characteristics of data by showing them as hierarchically connected nodes. It takes the data at the lowest level of the hierarchy as input, visualizes the data as a three-dimensional \landscape", and presents the data as clustered, hierarchical blocks (node) and bars through which part or all of the data set can be dynamically navigated and viewed. In addition, several researchers have also studied extensively the decision tree visualization, like the interactive visualization of decision tree construction [5] and the interactive construction of two- and threedimensional decision trees [1]. Items association rules visualization and neural networks visualization are discussed in [1], while the numerical association rules visualization methods can be found in [15, 19]. The third component is to discretize numerical (continuous) attributes. The CViz system provides three approaches: equi-length, bin-packing based method [19], and entropy-based method EDA-DB [4]. The user can select the method that he/she likes, or select all of them for dierent learning runs for comparison of the learning results. The discretized attributes are visualized again where an attribute interval corresponds to a point on the attribute coordinate. Figures 4 and 10 show the discretized parallel coordinates. The CViz system draw the classication rules on the parallel coordinates as rule polygons. 3. For the pattern discovery, a rule induction algorithm, ELEM2, is used to learn classication rules [3]. ELEM2 conducts a general-to-specic heuristic search over a hypothesis space to sequentially generate a disjunctive set of propositional rules for each class. It uses a sequential covering learning strategy to reduce the problem of learning a disjunctive set of rules to a sequence of simpler problems, each requiring that a single conjunctive rule be learned that covers a subset of positive examples. THE CVIZ SYSTEM We develop an interactive system, CViz, for visualizing the process of classication rule induction according to the RuleViz model. The CViz system is based on the parallel coordinates technique, a kind of geometric techniques [24]. CViz uses interactive visualization techniques to visualize the original data, help the user clean and preprocess the data, and interpret the rules discovered. CViz also conducts numerical attribute discretization and rule induction to learn classication rules based on the training data set. CViz consists of ve components, original data visualization, data reduction, numerical attribute discretization, classication rule induction and classication rule visualization. In this section, we briey introduce the implementation of the CViz system and demonstrate our experiment results with the UCI data sets and articial data set. 3.1 The last component is the visualization of the discovered classication rules. The nal rules are displayed as colored strips (polygons), and the rule accuracy and rule quality are used to render the rule polygons. Each rule is illustrated by a subset of coordinates with corresponding values and/or intervals. The coordinates in the subset are connected together through the values or intervals. The rule accuracy and quality values are used to render the rule strips. CViz Implementation If many rules are obtained and visualized on a same picture, it is hard to distinguish them. CViz allows user to specify a class label to view all rules that have this class label as the decision value. Additionally, the rule prediction accuracy and rule quality are used to control the display of the discovered rules to avoid clutter. Two sliders are used to set the range of rule accuracy, and one slider is used to specify the rule quality threshold. If these control sliders are set, only can the rules with accuracy in the accuracy range and quality higher than the quality threshold be displayed. As the sliders move, the displayed rules change. The CViz system is implemented with Visual C++ 6.0 in terms of the RuleViz model. The rst step to visually discover knowledge is visualizing the original data. CViz assumes the training data are represented as n-ary tuples. n equidistant axes are created to be parallel to one of the screen axes, say Y-axis, and correspond to the attributes. The axes are scaled to the range of the corresponding attributes and normalized, if necessary. Every tuple corresponds to a polyline which intersects each of the axes at the point that corresponds to the value for the attribute [22]. Figures 3 and 9 shows the eects of using parallel coordinates technique to visualize multidimensional data sets. Each coordinate is interpreted further with a list of values for a categorical attribute, while a numerical attribute can be attached with the minimum and maximum, even mean, variance, etc. As the algorithm proceeds, numerical intervals can be added, including the start point and the end point for each interval and the number of total intervals once the numerical attribute is discretized. 3.2 Rule Visualization The basic idea of visualizing classication rules in the CViz system is to represent a rule as a strip, called rule polygon, which covers the area that connects the corresponding attribute values. A classication rule induced by the ELEM2 algorithm is a logical statement that consists of two parts: condition part and decision part. The decision part is simply the class label. The condition part is composed of a set of simple relations, each of which has the format of < a; r; v > as above, where r 2 f=; ! =g and v is a set of values of a if a is a categorical attribute, or r 2 f<=; >; 2g and v is a single value for <= and > or a pair of values for 2 if a is a numerical attribute. The following is the explanation for each relational operator. The second component is to reduce the original data, if necessary. Once the original data is visualized on the parallel coordinates, the user can get a rough idea about the data distribution so that data cleaning might be done. The data reduction involves two aspects: horizontal and vertical reduction. An attribute that the user thinks irrelevant to 248 Procedure: DrawRule for classication rule visualization preCond = rst condition; repeat postCond = next condition; for each or component of preCond do if preCond corresponds to a categorical attribute then compute the points p0 ; p1 according to a specic value; else compute the points p0 ; p1 according to an interval; for each or component of postCond do if postCond corresponds to a categorical attribute then compute the points p2 ; p3 according to a specic value; else compute the points p2 ; p3 according to an interval; draw a small polygon with four points p0 ; p1 ; p2 ; and p3 ; preCond = postCond; until all conditions are processed; draw the small polygon from last condition to the decision attribute. Figure 2: Procedure for drawing rule polygons a = fx1 ; x2 ; : : : ; xn g: attribute a has any values of x1 ; x2 ; : : : , or xn , where n >= 1; a! = fx1 ; x2 ; : : : ; xn g: attribute a has any possible values except x1 ; x2 ; : : : , and xn , where n >= 1. a <= x: attribute a has any values between the minimum and x; a > x: attribute a has any values between x and the maximum; a 2 (x1 ; x2 ]: and x2 . corresponding to the rst condition to the decision coordinate in terms of the condition part and is depicted in Figure 2. To distinguish between positive conditions represented with relational operator = and negative conditions represented with relational operator !=, the positive condition is drawn with forward hatch, while the negative condition is drawn with backward hatch. For example, in Figure 11, we have a strip from Color = red to Class = good with backward hatch, which indicates a rule < Color! = red >) Class = good: attribute a has any values between x1 In addition, the rule accuracy and the rule normalized quality are used to calculate the color of the rule polygon. The more accurate the rule, the redder the polygon, while the lower the rule quality, the more bright the polygon, For example, the following is a classication rule obtained in our experiment with an articial data set, and is visualized in Figure 11: 3.3 CViz Experiment CViz has been experimented with several data sets from UCI Repository of Machine Learning Databases [28], including IRIS, Monk's, etc. and also with articial data sets. The IRIS data are represented by only continuous attributes except the class label, while the Monk data set consists of only discrete attributes. In the articial data set, we design it contains both numerical and discrete attributes. < Age 2 (30; 60] >< Score 2 (1:5; 3:5] > < Color = fred; yellowg >) Class = bad: This rule means that if numerical attributes Age and Score have values between 30 and 60 and between 1:5 and 3:5, respectively, and categorical attribute Color has value red or yellow, then the instance will be in class bad. The corresponding rule polygon consists of two polygons which are enclosed by the points that correspond to value red and yellow on coordinate Color, respectively, the points correspond to 30 and 60 on coordinate Age, the points that correspond to 1:5 and 3:5 on coordinate Score, and the point corresponding to class bad on the decision coordinate. Actually, above rule can be logically decomposed into two equivalent rules, corresponding to < Color = red > and < Color = yellow >, respectively. 3.3.1 Experiment with UCI data sets The IRIS data set contains three classes of iris plants with 50 objects belonging to each class. One of the iris types is linearly separable from the other two while the other two are not linearly separable from each other. Each object is described by four numerical attributes, sepal-length, sepalwidth, petal-length, and petal-width. Three classes are IRIS Setosa, IRIS Versicolor, and IRIS Virginica. Figures 3 through 7 demonstrate the experimental results with the IRIS data set. Figure 3 is the original data visualization, while Figure 4 is the visualization of discretized The rule polygon is constructed by a rule polygon generating procedure, which draws the rule strip from the coordinate 249 Figure 3: The original IRIS data Figure 5: The rules for the Figure 4: The IRIS attributes discretization with EDA-DB method IRIS data Figure 6: The rules for IRIS class 2 numerical attributes with EDA-DB method. From Figure 4, we see that four numerical attributes are discretized into dierent number of intervals with dierent width. For example, attribute petal-width is discretized into eight disjoint intervals, and the boundary points are 0:0, 0:6, 1:3, 1:4, 1:5, 1:6, 1:7, 1:8, and 2:5, respectively. Figure 5 shows all rules that ELEM2 learns from the IRIS data sets, which involves eight classication rules, including four rules for class 3, three rules for class 2, and one rule for class 1, respectively. Figure 6 illustrates all rules for class 2 discovered from the IRIS data set, and the rules for class 1 and class 3 discovered from the IRIS data set have the similar display form. Figure 7 shows the rules for class 2 with accuracy greater than 90%, and quality greater than 85%. From Figure 7, it is easy to see we obtain two rules that satisfy the accuracy and quality specication through sliders, which are < sepal length > 5:8 >< sepal < petal length <= 4:8 > Figure 7: The rules for greater than 85% width > 3:0 > ) Class = 2: 250 IRIS class 2 with quality Figure 8: The rules for the Monk-1 data and < petal < petal Figure 9: The original articial data length 2 (1:9; 4:8] > width < 1:6 > ) Class = 2: We also experiment the CViz system with the Monk-1 data set, which consists of a set of robots described by six attributes: head shapes, body shapes, facial expressions, objects being held, jacket colors, and whether or not the robot is wearing a tie. The last attribute is a decision attribute (whether a robot belongs to a positive or negative class). This data set contains 432 examples. The rules obtained are illustrated in Figure 8, from which we can see there are nine classication rules where ve rules are for class 0 (negative class) and four rules for class 1 (positive class). 3.3.2 Experiment with artificial data An articial data set is designed to involve two numerical attributes Age and Score, one condition categorical attribute Color, and one decision categorical attribute Class. Age is of integer type and has range [0; 90]. Score is of real-valued type and has range [0:0; 5:0]. Color has four discrete values: red, blue, yellow, and green. The decision attribute has two class labels: good and bad. The relationship between the class labels and the condition attributes in the data set is designed as follows: Figure 10: The numerical attributes discretization of the articial data set Age<=30 or Age>60, Score<=1.5 or Score>3.5, Color!=(red or yellow). Figure 12 shows, we have only two rules with accuracy between 90% and 95% and quality greater than 90%, one for class good, one for class bad. While Figure 13 illustrates the rules for class good with accuracy greater than 95%. If (30<Age<=60) and (1.5<Score<=3.5) and (Color=red or yellow) Then Class=bad Otherwise Class=good. 4. CONCLUDING REMARKS RuleViz is an interactive model for visualizing the process of knowledge discovery in databases. This model consists of ve components, original data visualization, visual data reduction, visual data preprocess, pattern discovery, and pattern visualization. The basic idea is to use visualization techniques to draw the original data and get insight into the data distribution, intuitively limit the domain of data by interaction, then discover rules from the reduced data, and nally visualize the resulting knowledge. Two aspects are addressed. One is the interaction between the human Figures 9 through 13 illustrate the experiment with the articial data set. Figure 9 visualizes the original articial data, and Figure 10 is the result of discretizing the numerical attributes Age and Score with EDA-DB method [4]. Figure 11 illustrates all discovered rules, including one rule for class bad and three rules for class good, which are the same as our designation and correspond to the following three conditions: 251 and the machine, and the other is the visualization of each component. The interaction helps the KDD system navigate through the enormous search spaces and recognize the intentions of the user, and the visualization of the KDD process helps users gain better insight into the multidimensional data, understand the intermediate results, and interpret the discovered patterns. The information visualization techniques can be used for visualizing each component, but new techniques for interactive visualization should be developed. According to the RuleViz model, an interactive visualization system, CViz, for visualizing the process of classication rule induction is developed. In our implementation, we emphasize the human-machine interaction, since we believe that interactive visualization plays an important role in data mining and machine learning to guide the process of discovering knowledge. The experimental results with the UCI data, and articial data have also demonstrated that it is useful for users to understand the relationships among data and to concentrate on the interesting data to discover knowledge. Figure 11: All rules discovered with articial data set Enhancing the CViz system, developing other systems for visually discovering other types of knowledge, and nally integrating these systems under the RuleViz model constitute our next work. The capabilities of CViz will be improved with other visualization techniques and expanded to provide more interactive tools for the data preprocess, such as interactively drawing clusters so that the clustering problem can be transformed to classication problems. We are also adapting the CViz system to other rule-based learning systems so that CViz contains many rule induction algorithms for dierent domains. Other interactive systems based on the RuleViz model are being developed to visualize the process of mining association rules, evolution rules, etc. 5. ACKNOWLEDGMENTS 6. REFERENCES The authors are members of the Institute for Robotics and Intelligent Systems (IRIS) and wish to acknowledge the support of the Networks of Centers of Excellence Program of the Government of Canada, the Natural Sciences and Engineering Research Council, and the participation of PRECARN Associates Inc. Figure 12: The rules from articial data with accuracy between 90% and 95% [1] P. Adriaans and D. Zantinge. Data Mining. Addison-Wesley, 1997. [2] C. Ahlberg. Spotre: An information exploration environment. ACM SIGMOD Record, 25(4):25{29, 1996. [3] A. An and N. Cercone. Elem2: A learning system for more accurate classications. In Proc. of the 12th Biennial Conference of the Canadian Society for Computational Studies of Intelligence, pages 426{441, 1998. [4] A. An and N. Cercone. Discretization of continuous attributes for learning classication rules. In Proc. of the 3rd Pacic-Asia Conference on Knowledge Discovery in Databases, pages 509{514, 1999. Figure 13: The rules for the good class with accuracy between 95% and 100% 252 [5] M. Ankerst, C. Elsen, M. Ester, and H. P. Kriegel. Visual classication: An interactive approach to decision tree construction. In Proc. of KDD-99, pages 392{397, 1999. [18] J. Han and N. Cercone. Dviz: A system for visualizing data mining. In Proc. of the 3rd Pacic-Asia Conference on Knowledge Discovery in Databases, pages 390{399, 1999. [6] D. Barbara, W. DuMouchel, C. Faloutsos, P. J. Haas, J. M. Hellerstein, Y. Ioannidis, H. V. Jagadish, T. Johnson, V. P. R. Ng, K. A. Ross, and K. C. Sevcik. The new jersey data reduction report. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering, pages 3{45, 1997. [19] J. Han and N. Cercone. Aviz: A visualization system for discovering numerical association rules. In Proc. of the 4th Pacic-Asia Conference on Knowledge Discovery in Databases, pages 269{280, 2000. [20] J. Han and N. Cercone. Cviz: A visualization system for rule induction. In Proc. of the 13th Biennial Conference of the Canadian Society for Computational Studies of Intelligence, pages 214{226, 2000. [7] S. Berchtold, H. V. Jagadish, and K. A. Ross. Independence diagrams: A technique for visual data mining. In Proc. KDD-98, pages 139{143, 1998. [21] J. Han and N. Cercone. Typical example selection for rule induction. In Proc. of the 13th Biennial Conference of the Canadian Society for Computational Studies of Intelligence, pages 347{356, 2000. [8] R. J. Brachman and T. Anand. The process of knowledge discovery in databases: A human-centered approach. In Advances in Knowledge Discovery and Data Mining, eds. by U. M. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy, pages 37{58. AAAI Press / The MIT Press, 1996. [22] A. Inselberg and B. Dimsdale. Parallel coordinates: A tool for visualizing multi-dimensional geometry. In Proc of IEEE Visualization Conference, pages 361{370, 1990. [9] C. Brunk, J. Kelly, and R. Kohavi. Mineset: An integrated system for data mining. In Proc. of KDD-97, pages 135{138, 1997. [23] D. A. Keim and H. P. Kriegel. Visdb: Database exploration using multidimensional visualization. IEEE Computer Graphics and Applications, pages 40{49, 1994. [10] K. Cios, W. Pedrycz, and R. Swiniarski. Data Mining Methods for Knowledge Discovery. Kluwer Academic Publishers, 1998. [24] D. A. Keim and H. P. Kriegel. Visualization techniques for mining large databases: A comparison. Transactions on Knowledge and Data Engineering, 8(6):923{938, 1996. [11] J. V. Curtis and J. A. Konstan. Interactive visualization of periodic data. In Proc. of the ACM Symp. of User Interface Software and Visualization, pages 29{38, 1998. [25] H. Liu and H. Motoda. Feature Extraction Construction and Selection: A Data Mining Perspective. Kluwer Academic Publishers, 1998. [12] M. Derthick, J. Kolojejchick, and S. F. Roth. An interactive visualization environment for data exploration. In Proc. of KDD-97, pages 2{9, 1997. [26] R. J. Miller and Y. Yang. Association rules over interval data. In Proc. of the ACM SIGMOD International Conference on Management of Data, pages 452{461, 1997. [13] U. M. Fayyad, G. Piatetsky-Shapiro, and P. Smyth. From data mining to knowledge discovery: An overview. In Advances in Knowledge Discovery and Data Mining, eds. by U. M. Fayyad and G. Piatetsky-Shapiro and P. Smyth, and R. Uthurusamy, pages 1{36. AAAI Press / The MIT Press, 1996. [27] T. M. Mitchell. Machine Learning. McGraw-Hill, 1997. [28] P. M. Murphy and D. W. Aho. Uci repository of machine learning databases. http://www.ics.uci.edu/mlearn/MLRepository.html, ~ 1994. [14] W. J. Frawley, G. Piatetsky-Shapiro, and C. J. Matheus. Knowledge discovery in databases: An overview. In Knowledge Discovery in Databases, eds. by G. Piatetsky-Shapiro and W. J. Frawley, pages 1{27. AAAI Press / The MIT Press, 1991. [29] G. Piatetsky-Shapiro. Discovery, analysis, and presentation of strong rules. In Knowledge Discovery in Databases, eds. by G. Piatetsky-Shapiro and W. J. Frawley, pages 229{260, 1991. [15] T. Fukuda, Y. Morimoto, S. Morishita, and T. Tokuyama. Data mining using two-dimensional optimized association rules: Scheme, algorithms, and visualization. In Proc. of the ACM SIGMOD International Conference on Management of Data, pages 13{24, 1996. [30] J. R. Quinlan. C4.5: Programs for Machine Learning. CA: Morgan Kaufmann, 1993. [16] A. D. Gordon. Classication. Chapman and Hall, 1981. [17] R. Groth. Data Mining: A Hands-on Approach for Business Professionals. Prentice Hall PTR, 1998. 253